1 /* 2 * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "libadt/vectset.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/convertnode.hpp" 34 #include "opto/locknode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/macro.hpp" 37 #include "opto/memnode.hpp" 38 #include "opto/narrowptrnode.hpp" 39 #include "opto/node.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/phaseX.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/subnode.hpp" 45 #include "opto/type.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 49 // 50 // Replace any references to "oldref" in inputs to "use" with "newref". 51 // Returns the number of replacements made. 52 // 53 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 54 int nreplacements = 0; 55 uint req = use->req(); 56 for (uint j = 0; j < use->len(); j++) { 57 Node *uin = use->in(j); 58 if (uin == oldref) { 59 if (j < req) 60 use->set_req(j, newref); 61 else 62 use->set_prec(j, newref); 63 nreplacements++; 64 } else if (j >= req && uin == NULL) { 65 break; 66 } 67 } 68 return nreplacements; 69 } 70 71 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 72 // Copy debug information and adjust JVMState information 73 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 74 uint new_dbg_start = newcall->tf()->domain()->cnt(); 75 int jvms_adj = new_dbg_start - old_dbg_start; 76 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 77 78 // SafePointScalarObject node could be referenced several times in debug info. 79 // Use Dict to record cloned nodes. 80 Dict* sosn_map = new Dict(cmpkey,hashkey); 81 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 82 Node* old_in = oldcall->in(i); 83 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 84 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 85 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 86 uint old_unique = C->unique(); 87 Node* new_in = old_sosn->clone(sosn_map); 88 if (old_unique != C->unique()) { // New node? 89 new_in->set_req(0, C->root()); // reset control edge 90 new_in = transform_later(new_in); // Register new node. 91 } 92 old_in = new_in; 93 } 94 newcall->add_req(old_in); 95 } 96 97 newcall->set_jvms(oldcall->jvms()); 98 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 99 jvms->set_map(newcall); 100 jvms->set_locoff(jvms->locoff()+jvms_adj); 101 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 102 jvms->set_monoff(jvms->monoff()+jvms_adj); 103 jvms->set_scloff(jvms->scloff()+jvms_adj); 104 jvms->set_endoff(jvms->endoff()+jvms_adj); 105 } 106 } 107 108 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 109 Node* cmp; 110 if (mask != 0) { 111 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 112 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 113 } else { 114 cmp = word; 115 } 116 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 117 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 118 transform_later(iff); 119 120 // Fast path taken. 121 Node *fast_taken = transform_later(new IfFalseNode(iff)); 122 123 // Fast path not-taken, i.e. slow path 124 Node *slow_taken = transform_later(new IfTrueNode(iff)); 125 126 if (return_fast_path) { 127 region->init_req(edge, slow_taken); // Capture slow-control 128 return fast_taken; 129 } else { 130 region->init_req(edge, fast_taken); // Capture fast-control 131 return slow_taken; 132 } 133 } 134 135 //--------------------copy_predefined_input_for_runtime_call-------------------- 136 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 137 // Set fixed predefined input arguments 138 call->init_req( TypeFunc::Control, ctrl ); 139 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 140 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 141 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 142 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 143 } 144 145 //------------------------------make_slow_call--------------------------------- 146 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) { 147 148 // Slow-path call 149 CallNode *call = leaf_name 150 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 151 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 152 153 // Slow path call has no side-effects, uses few values 154 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 155 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 156 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 157 copy_call_debug_info(oldcall, call); 158 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 159 _igvn.replace_node(oldcall, call); 160 transform_later(call); 161 162 return call; 163 } 164 165 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 166 _fallthroughproj = NULL; 167 _fallthroughcatchproj = NULL; 168 _ioproj_fallthrough = NULL; 169 _ioproj_catchall = NULL; 170 _catchallcatchproj = NULL; 171 _memproj_fallthrough = NULL; 172 _memproj_catchall = NULL; 173 _resproj = NULL; 174 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 175 ProjNode *pn = call->fast_out(i)->as_Proj(); 176 switch (pn->_con) { 177 case TypeFunc::Control: 178 { 179 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 180 _fallthroughproj = pn; 181 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 182 const Node *cn = pn->fast_out(j); 183 if (cn->is_Catch()) { 184 ProjNode *cpn = NULL; 185 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 186 cpn = cn->fast_out(k)->as_Proj(); 187 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 188 if (cpn->_con == CatchProjNode::fall_through_index) 189 _fallthroughcatchproj = cpn; 190 else { 191 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 192 _catchallcatchproj = cpn; 193 } 194 } 195 } 196 break; 197 } 198 case TypeFunc::I_O: 199 if (pn->_is_io_use) 200 _ioproj_catchall = pn; 201 else 202 _ioproj_fallthrough = pn; 203 break; 204 case TypeFunc::Memory: 205 if (pn->_is_io_use) 206 _memproj_catchall = pn; 207 else 208 _memproj_fallthrough = pn; 209 break; 210 case TypeFunc::Parms: 211 _resproj = pn; 212 break; 213 default: 214 assert(false, "unexpected projection from allocation node."); 215 } 216 } 217 218 } 219 220 // Eliminate a card mark sequence. p2x is a ConvP2XNode 221 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 222 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 223 if (!UseG1GC) { 224 // vanilla/CMS post barrier 225 Node *shift = p2x->unique_out(); 226 Node *addp = shift->unique_out(); 227 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 228 Node *mem = addp->last_out(j); 229 if (UseCondCardMark && mem->is_Load()) { 230 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 231 // The load is checking if the card has been written so 232 // replace it with zero to fold the test. 233 _igvn.replace_node(mem, intcon(0)); 234 continue; 235 } 236 assert(mem->is_Store(), "store required"); 237 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 238 } 239 } else { 240 // G1 pre/post barriers 241 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 242 // It could be only one user, URShift node, in Object.clone() instrinsic 243 // but the new allocation is passed to arraycopy stub and it could not 244 // be scalar replaced. So we don't check the case. 245 246 // An other case of only one user (Xor) is when the value check for NULL 247 // in G1 post barrier is folded after CCP so the code which used URShift 248 // is removed. 249 250 // Take Region node before eliminating post barrier since it also 251 // eliminates CastP2X node when it has only one user. 252 Node* this_region = p2x->in(0); 253 assert(this_region != NULL, ""); 254 255 // Remove G1 post barrier. 256 257 // Search for CastP2X->Xor->URShift->Cmp path which 258 // checks if the store done to a different from the value's region. 259 // And replace Cmp with #0 (false) to collapse G1 post barrier. 260 Node* xorx = NULL; 261 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) { 262 Node* u = p2x->fast_out(i); 263 if (u->Opcode() == Op_XorX) { 264 xorx = u; 265 break; 266 } 267 } 268 assert(xorx != NULL, "missing G1 post barrier"); 269 Node* shift = xorx->unique_out(); 270 Node* cmpx = shift->unique_out(); 271 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 272 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 273 "missing region check in G1 post barrier"); 274 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 275 276 // Remove G1 pre barrier. 277 278 // Search "if (marking != 0)" check and set it to "false". 279 // There is no G1 pre barrier if previous stored value is NULL 280 // (for example, after initialization). 281 if (this_region->is_Region() && this_region->req() == 3) { 282 int ind = 1; 283 if (!this_region->in(ind)->is_IfFalse()) { 284 ind = 2; 285 } 286 if (this_region->in(ind)->is_IfFalse()) { 287 Node* bol = this_region->in(ind)->in(0)->in(1); 288 assert(bol->is_Bool(), ""); 289 cmpx = bol->in(1); 290 if (bol->as_Bool()->_test._test == BoolTest::ne && 291 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 292 cmpx->in(1)->is_Load()) { 293 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 294 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 295 PtrQueue::byte_offset_of_active()); 296 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 297 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 298 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 299 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 300 } 301 } 302 } 303 } 304 // Now CastP2X can be removed since it is used only on dead path 305 // which currently still alive until igvn optimize it. 306 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 307 _igvn.replace_node(p2x, top()); 308 } 309 } 310 311 // Search for a memory operation for the specified memory slice. 312 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 313 Node *orig_mem = mem; 314 Node *alloc_mem = alloc->in(TypeFunc::Memory); 315 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 316 while (true) { 317 if (mem == alloc_mem || mem == start_mem ) { 318 return mem; // hit one of our sentinels 319 } else if (mem->is_MergeMem()) { 320 mem = mem->as_MergeMem()->memory_at(alias_idx); 321 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 322 Node *in = mem->in(0); 323 // we can safely skip over safepoints, calls, locks and membars because we 324 // already know that the object is safe to eliminate. 325 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 326 return in; 327 } else if (in->is_Call()) { 328 CallNode *call = in->as_Call(); 329 if (!call->may_modify(tinst, phase)) { 330 mem = call->in(TypeFunc::Memory); 331 } 332 mem = in->in(TypeFunc::Memory); 333 } else if (in->is_MemBar()) { 334 mem = in->in(TypeFunc::Memory); 335 } else { 336 assert(false, "unexpected projection"); 337 } 338 } else if (mem->is_Store()) { 339 const TypePtr* atype = mem->as_Store()->adr_type(); 340 int adr_idx = Compile::current()->get_alias_index(atype); 341 if (adr_idx == alias_idx) { 342 assert(atype->isa_oopptr(), "address type must be oopptr"); 343 int adr_offset = atype->offset(); 344 uint adr_iid = atype->is_oopptr()->instance_id(); 345 // Array elements references have the same alias_idx 346 // but different offset and different instance_id. 347 if (adr_offset == offset && adr_iid == alloc->_idx) 348 return mem; 349 } else { 350 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 351 } 352 mem = mem->in(MemNode::Memory); 353 } else if (mem->is_ClearArray()) { 354 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 355 // Can not bypass initialization of the instance 356 // we are looking. 357 debug_only(intptr_t offset;) 358 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 359 InitializeNode* init = alloc->as_Allocate()->initialization(); 360 // We are looking for stored value, return Initialize node 361 // or memory edge from Allocate node. 362 if (init != NULL) 363 return init; 364 else 365 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 366 } 367 // Otherwise skip it (the call updated 'mem' value). 368 } else if (mem->Opcode() == Op_SCMemProj) { 369 mem = mem->in(0); 370 Node* adr = NULL; 371 if (mem->is_LoadStore()) { 372 adr = mem->in(MemNode::Address); 373 } else { 374 assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); 375 adr = mem->in(3); // Destination array 376 } 377 const TypePtr* atype = adr->bottom_type()->is_ptr(); 378 int adr_idx = Compile::current()->get_alias_index(atype); 379 if (adr_idx == alias_idx) { 380 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 381 return NULL; 382 } 383 mem = mem->in(MemNode::Memory); 384 } else { 385 return mem; 386 } 387 assert(mem != orig_mem, "dead memory loop"); 388 } 389 } 390 391 // 392 // Given a Memory Phi, compute a value Phi containing the values from stores 393 // on the input paths. 394 // Note: this function is recursive, its depth is limied by the "level" argument 395 // Returns the computed Phi, or NULL if it cannot compute it. 396 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) { 397 assert(mem->is_Phi(), "sanity"); 398 int alias_idx = C->get_alias_index(adr_t); 399 int offset = adr_t->offset(); 400 int instance_id = adr_t->instance_id(); 401 402 // Check if an appropriate value phi already exists. 403 Node* region = mem->in(0); 404 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 405 Node* phi = region->fast_out(k); 406 if (phi->is_Phi() && phi != mem && 407 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 408 return phi; 409 } 410 } 411 // Check if an appropriate new value phi already exists. 412 Node* new_phi = value_phis->find(mem->_idx); 413 if (new_phi != NULL) 414 return new_phi; 415 416 if (level <= 0) { 417 return NULL; // Give up: phi tree too deep 418 } 419 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 420 Node *alloc_mem = alloc->in(TypeFunc::Memory); 421 422 uint length = mem->req(); 423 GrowableArray <Node *> values(length, length, NULL, false); 424 425 // create a new Phi for the value 426 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 427 transform_later(phi); 428 value_phis->push(phi, mem->_idx); 429 430 for (uint j = 1; j < length; j++) { 431 Node *in = mem->in(j); 432 if (in == NULL || in->is_top()) { 433 values.at_put(j, in); 434 } else { 435 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 436 if (val == start_mem || val == alloc_mem) { 437 // hit a sentinel, return appropriate 0 value 438 values.at_put(j, _igvn.zerocon(ft)); 439 continue; 440 } 441 if (val->is_Initialize()) { 442 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 443 } 444 if (val == NULL) { 445 return NULL; // can't find a value on this path 446 } 447 if (val == mem) { 448 values.at_put(j, mem); 449 } else if (val->is_Store()) { 450 values.at_put(j, val->in(MemNode::ValueIn)); 451 } else if(val->is_Proj() && val->in(0) == alloc) { 452 values.at_put(j, _igvn.zerocon(ft)); 453 } else if (val->is_Phi()) { 454 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 455 if (val == NULL) { 456 return NULL; 457 } 458 values.at_put(j, val); 459 } else if (val->Opcode() == Op_SCMemProj) { 460 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity"); 461 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 462 return NULL; 463 } else { 464 #ifdef ASSERT 465 val->dump(); 466 assert(false, "unknown node on this path"); 467 #endif 468 return NULL; // unknown node on this path 469 } 470 } 471 } 472 // Set Phi's inputs 473 for (uint j = 1; j < length; j++) { 474 if (values.at(j) == mem) { 475 phi->init_req(j, phi); 476 } else { 477 phi->init_req(j, values.at(j)); 478 } 479 } 480 return phi; 481 } 482 483 // Search the last value stored into the object's field. 484 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 485 assert(adr_t->is_known_instance_field(), "instance required"); 486 int instance_id = adr_t->instance_id(); 487 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 488 489 int alias_idx = C->get_alias_index(adr_t); 490 int offset = adr_t->offset(); 491 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 492 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 493 Node *alloc_mem = alloc->in(TypeFunc::Memory); 494 Arena *a = Thread::current()->resource_area(); 495 VectorSet visited(a); 496 497 498 bool done = sfpt_mem == alloc_mem; 499 Node *mem = sfpt_mem; 500 while (!done) { 501 if (visited.test_set(mem->_idx)) { 502 return NULL; // found a loop, give up 503 } 504 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 505 if (mem == start_mem || mem == alloc_mem) { 506 done = true; // hit a sentinel, return appropriate 0 value 507 } else if (mem->is_Initialize()) { 508 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 509 if (mem == NULL) { 510 done = true; // Something go wrong. 511 } else if (mem->is_Store()) { 512 const TypePtr* atype = mem->as_Store()->adr_type(); 513 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 514 done = true; 515 } 516 } else if (mem->is_Store()) { 517 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 518 assert(atype != NULL, "address type must be oopptr"); 519 assert(C->get_alias_index(atype) == alias_idx && 520 atype->is_known_instance_field() && atype->offset() == offset && 521 atype->instance_id() == instance_id, "store is correct memory slice"); 522 done = true; 523 } else if (mem->is_Phi()) { 524 // try to find a phi's unique input 525 Node *unique_input = NULL; 526 Node *top = C->top(); 527 for (uint i = 1; i < mem->req(); i++) { 528 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 529 if (n == NULL || n == top || n == mem) { 530 continue; 531 } else if (unique_input == NULL) { 532 unique_input = n; 533 } else if (unique_input != n) { 534 unique_input = top; 535 break; 536 } 537 } 538 if (unique_input != NULL && unique_input != top) { 539 mem = unique_input; 540 } else { 541 done = true; 542 } 543 } else { 544 assert(false, "unexpected node"); 545 } 546 } 547 if (mem != NULL) { 548 if (mem == start_mem || mem == alloc_mem) { 549 // hit a sentinel, return appropriate 0 value 550 return _igvn.zerocon(ft); 551 } else if (mem->is_Store()) { 552 return mem->in(MemNode::ValueIn); 553 } else if (mem->is_Phi()) { 554 // attempt to produce a Phi reflecting the values on the input paths of the Phi 555 Node_Stack value_phis(a, 8); 556 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 557 if (phi != NULL) { 558 return phi; 559 } else { 560 // Kill all new Phis 561 while(value_phis.is_nonempty()) { 562 Node* n = value_phis.node(); 563 _igvn.replace_node(n, C->top()); 564 value_phis.pop(); 565 } 566 } 567 } 568 } 569 // Something go wrong. 570 return NULL; 571 } 572 573 // Check the possibility of scalar replacement. 574 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 575 // Scan the uses of the allocation to check for anything that would 576 // prevent us from eliminating it. 577 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 578 DEBUG_ONLY( Node* disq_node = NULL; ) 579 bool can_eliminate = true; 580 581 Node* res = alloc->result_cast(); 582 const TypeOopPtr* res_type = NULL; 583 if (res == NULL) { 584 // All users were eliminated. 585 } else if (!res->is_CheckCastPP()) { 586 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 587 can_eliminate = false; 588 } else { 589 res_type = _igvn.type(res)->isa_oopptr(); 590 if (res_type == NULL) { 591 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 592 can_eliminate = false; 593 } else if (res_type->isa_aryptr()) { 594 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 595 if (length < 0) { 596 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 597 can_eliminate = false; 598 } 599 } 600 } 601 602 if (can_eliminate && res != NULL) { 603 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 604 j < jmax && can_eliminate; j++) { 605 Node* use = res->fast_out(j); 606 607 if (use->is_AddP()) { 608 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 609 int offset = addp_type->offset(); 610 611 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 612 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 613 can_eliminate = false; 614 break; 615 } 616 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 617 k < kmax && can_eliminate; k++) { 618 Node* n = use->fast_out(k); 619 if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 620 DEBUG_ONLY(disq_node = n;) 621 if (n->is_Load() || n->is_LoadStore()) { 622 NOT_PRODUCT(fail_eliminate = "Field load";) 623 } else { 624 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 625 } 626 can_eliminate = false; 627 } 628 } 629 } else if (use->is_SafePoint()) { 630 SafePointNode* sfpt = use->as_SafePoint(); 631 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 632 // Object is passed as argument. 633 DEBUG_ONLY(disq_node = use;) 634 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 635 can_eliminate = false; 636 } 637 Node* sfptMem = sfpt->memory(); 638 if (sfptMem == NULL || sfptMem->is_top()) { 639 DEBUG_ONLY(disq_node = use;) 640 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 641 can_eliminate = false; 642 } else { 643 safepoints.append_if_missing(sfpt); 644 } 645 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 646 if (use->is_Phi()) { 647 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 648 NOT_PRODUCT(fail_eliminate = "Object is return value";) 649 } else { 650 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 651 } 652 DEBUG_ONLY(disq_node = use;) 653 } else { 654 if (use->Opcode() == Op_Return) { 655 NOT_PRODUCT(fail_eliminate = "Object is return value";) 656 }else { 657 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 658 } 659 DEBUG_ONLY(disq_node = use;) 660 } 661 can_eliminate = false; 662 } 663 } 664 } 665 666 #ifndef PRODUCT 667 if (PrintEliminateAllocations) { 668 if (can_eliminate) { 669 tty->print("Scalar "); 670 if (res == NULL) 671 alloc->dump(); 672 else 673 res->dump(); 674 } else if (alloc->_is_scalar_replaceable) { 675 tty->print("NotScalar (%s)", fail_eliminate); 676 if (res == NULL) 677 alloc->dump(); 678 else 679 res->dump(); 680 #ifdef ASSERT 681 if (disq_node != NULL) { 682 tty->print(" >>>> "); 683 disq_node->dump(); 684 } 685 #endif /*ASSERT*/ 686 } 687 } 688 #endif 689 return can_eliminate; 690 } 691 692 // Do scalar replacement. 693 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 694 GrowableArray <SafePointNode *> safepoints_done; 695 696 ciKlass* klass = NULL; 697 ciInstanceKlass* iklass = NULL; 698 int nfields = 0; 699 int array_base; 700 int element_size; 701 BasicType basic_elem_type; 702 ciType* elem_type; 703 704 Node* res = alloc->result_cast(); 705 const TypeOopPtr* res_type = NULL; 706 if (res != NULL) { // Could be NULL when there are no users 707 res_type = _igvn.type(res)->isa_oopptr(); 708 } 709 710 if (res != NULL) { 711 klass = res_type->klass(); 712 if (res_type->isa_instptr()) { 713 // find the fields of the class which will be needed for safepoint debug information 714 assert(klass->is_instance_klass(), "must be an instance klass."); 715 iklass = klass->as_instance_klass(); 716 nfields = iklass->nof_nonstatic_fields(); 717 } else { 718 // find the array's elements which will be needed for safepoint debug information 719 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 720 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 721 elem_type = klass->as_array_klass()->element_type(); 722 basic_elem_type = elem_type->basic_type(); 723 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 724 element_size = type2aelembytes(basic_elem_type); 725 } 726 } 727 // 728 // Process the safepoint uses 729 // 730 while (safepoints.length() > 0) { 731 SafePointNode* sfpt = safepoints.pop(); 732 Node* mem = sfpt->memory(); 733 assert(sfpt->jvms() != NULL, "missed JVMS"); 734 // Fields of scalar objs are referenced only at the end 735 // of regular debuginfo at the last (youngest) JVMS. 736 // Record relative start index. 737 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 738 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 739 #ifdef ASSERT 740 alloc, 741 #endif 742 first_ind, nfields); 743 sobj->init_req(0, C->root()); 744 transform_later(sobj); 745 746 // Scan object's fields adding an input to the safepoint for each field. 747 for (int j = 0; j < nfields; j++) { 748 intptr_t offset; 749 ciField* field = NULL; 750 if (iklass != NULL) { 751 field = iklass->nonstatic_field_at(j); 752 offset = field->offset(); 753 elem_type = field->type(); 754 basic_elem_type = field->layout_type(); 755 } else { 756 offset = array_base + j * (intptr_t)element_size; 757 } 758 759 const Type *field_type; 760 // The next code is taken from Parse::do_get_xxx(). 761 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 762 if (!elem_type->is_loaded()) { 763 field_type = TypeInstPtr::BOTTOM; 764 } else if (field != NULL && field->is_constant() && field->is_static()) { 765 // This can happen if the constant oop is non-perm. 766 ciObject* con = field->constant_value().as_object(); 767 // Do not "join" in the previous type; it doesn't add value, 768 // and may yield a vacuous result if the field is of interface type. 769 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 770 assert(field_type != NULL, "field singleton type must be consistent"); 771 } else { 772 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 773 } 774 if (UseCompressedOops) { 775 field_type = field_type->make_narrowoop(); 776 basic_elem_type = T_NARROWOOP; 777 } 778 } else { 779 field_type = Type::get_const_basic_type(basic_elem_type); 780 } 781 782 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 783 784 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 785 if (field_val == NULL) { 786 // We weren't able to find a value for this field, 787 // give up on eliminating this allocation. 788 789 // Remove any extra entries we added to the safepoint. 790 uint last = sfpt->req() - 1; 791 for (int k = 0; k < j; k++) { 792 sfpt->del_req(last--); 793 } 794 // rollback processed safepoints 795 while (safepoints_done.length() > 0) { 796 SafePointNode* sfpt_done = safepoints_done.pop(); 797 // remove any extra entries we added to the safepoint 798 last = sfpt_done->req() - 1; 799 for (int k = 0; k < nfields; k++) { 800 sfpt_done->del_req(last--); 801 } 802 JVMState *jvms = sfpt_done->jvms(); 803 jvms->set_endoff(sfpt_done->req()); 804 // Now make a pass over the debug information replacing any references 805 // to SafePointScalarObjectNode with the allocated object. 806 int start = jvms->debug_start(); 807 int end = jvms->debug_end(); 808 for (int i = start; i < end; i++) { 809 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 810 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 811 if (scobj->first_index(jvms) == sfpt_done->req() && 812 scobj->n_fields() == (uint)nfields) { 813 assert(scobj->alloc() == alloc, "sanity"); 814 sfpt_done->set_req(i, res); 815 } 816 } 817 } 818 } 819 #ifndef PRODUCT 820 if (PrintEliminateAllocations) { 821 if (field != NULL) { 822 tty->print("=== At SafePoint node %d can't find value of Field: ", 823 sfpt->_idx); 824 field->print(); 825 int field_idx = C->get_alias_index(field_addr_type); 826 tty->print(" (alias_idx=%d)", field_idx); 827 } else { // Array's element 828 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 829 sfpt->_idx, j); 830 } 831 tty->print(", which prevents elimination of: "); 832 if (res == NULL) 833 alloc->dump(); 834 else 835 res->dump(); 836 } 837 #endif 838 return false; 839 } 840 if (UseCompressedOops && field_type->isa_narrowoop()) { 841 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 842 // to be able scalar replace the allocation. 843 if (field_val->is_EncodeP()) { 844 field_val = field_val->in(1); 845 } else { 846 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 847 } 848 } 849 sfpt->add_req(field_val); 850 } 851 JVMState *jvms = sfpt->jvms(); 852 jvms->set_endoff(sfpt->req()); 853 // Now make a pass over the debug information replacing any references 854 // to the allocated object with "sobj" 855 int start = jvms->debug_start(); 856 int end = jvms->debug_end(); 857 sfpt->replace_edges_in_range(res, sobj, start, end); 858 safepoints_done.append_if_missing(sfpt); // keep it for rollback 859 } 860 return true; 861 } 862 863 // Process users of eliminated allocation. 864 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 865 Node* res = alloc->result_cast(); 866 if (res != NULL) { 867 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 868 Node *use = res->last_out(j); 869 uint oc1 = res->outcnt(); 870 871 if (use->is_AddP()) { 872 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 873 Node *n = use->last_out(k); 874 uint oc2 = use->outcnt(); 875 if (n->is_Store()) { 876 #ifdef ASSERT 877 // Verify that there is no dependent MemBarVolatile nodes, 878 // they should be removed during IGVN, see MemBarNode::Ideal(). 879 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 880 p < pmax; p++) { 881 Node* mb = n->fast_out(p); 882 assert(mb->is_Initialize() || !mb->is_MemBar() || 883 mb->req() <= MemBarNode::Precedent || 884 mb->in(MemBarNode::Precedent) != n, 885 "MemBarVolatile should be eliminated for non-escaping object"); 886 } 887 #endif 888 _igvn.replace_node(n, n->in(MemNode::Memory)); 889 } else { 890 eliminate_card_mark(n); 891 } 892 k -= (oc2 - use->outcnt()); 893 } 894 } else { 895 eliminate_card_mark(use); 896 } 897 j -= (oc1 - res->outcnt()); 898 } 899 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 900 _igvn.remove_dead_node(res); 901 } 902 903 // 904 // Process other users of allocation's projections 905 // 906 if (_resproj != NULL && _resproj->outcnt() != 0) { 907 // First disconnect stores captured by Initialize node. 908 // If Initialize node is eliminated first in the following code, 909 // it will kill such stores and DUIterator_Last will assert. 910 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 911 Node *use = _resproj->fast_out(j); 912 if (use->is_AddP()) { 913 // raw memory addresses used only by the initialization 914 _igvn.replace_node(use, C->top()); 915 --j; --jmax; 916 } 917 } 918 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 919 Node *use = _resproj->last_out(j); 920 uint oc1 = _resproj->outcnt(); 921 if (use->is_Initialize()) { 922 // Eliminate Initialize node. 923 InitializeNode *init = use->as_Initialize(); 924 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 925 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 926 if (ctrl_proj != NULL) { 927 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 928 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 929 } 930 Node *mem_proj = init->proj_out(TypeFunc::Memory); 931 if (mem_proj != NULL) { 932 Node *mem = init->in(TypeFunc::Memory); 933 #ifdef ASSERT 934 if (mem->is_MergeMem()) { 935 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 936 } else { 937 assert(mem == _memproj_fallthrough, "allocation memory projection"); 938 } 939 #endif 940 _igvn.replace_node(mem_proj, mem); 941 } 942 } else { 943 assert(false, "only Initialize or AddP expected"); 944 } 945 j -= (oc1 - _resproj->outcnt()); 946 } 947 } 948 if (_fallthroughcatchproj != NULL) { 949 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 950 } 951 if (_memproj_fallthrough != NULL) { 952 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 953 } 954 if (_memproj_catchall != NULL) { 955 _igvn.replace_node(_memproj_catchall, C->top()); 956 } 957 if (_ioproj_fallthrough != NULL) { 958 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 959 } 960 if (_ioproj_catchall != NULL) { 961 _igvn.replace_node(_ioproj_catchall, C->top()); 962 } 963 if (_catchallcatchproj != NULL) { 964 _igvn.replace_node(_catchallcatchproj, C->top()); 965 } 966 } 967 968 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 969 if (!EliminateAllocations || !alloc->_is_non_escaping) { 970 return false; 971 } 972 Node* klass = alloc->in(AllocateNode::KlassNode); 973 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 974 Node* res = alloc->result_cast(); 975 // Eliminate boxing allocations which are not used 976 // regardless scalar replacable status. 977 bool boxing_alloc = C->eliminate_boxing() && 978 tklass->klass()->is_instance_klass() && 979 tklass->klass()->as_instance_klass()->is_box_klass(); 980 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 981 return false; 982 } 983 984 extract_call_projections(alloc); 985 986 GrowableArray <SafePointNode *> safepoints; 987 if (!can_eliminate_allocation(alloc, safepoints)) { 988 return false; 989 } 990 991 if (!alloc->_is_scalar_replaceable) { 992 assert(res == NULL, "sanity"); 993 // We can only eliminate allocation if all debug info references 994 // are already replaced with SafePointScalarObject because 995 // we can't search for a fields value without instance_id. 996 if (safepoints.length() > 0) { 997 return false; 998 } 999 } 1000 1001 if (!scalar_replacement(alloc, safepoints)) { 1002 return false; 1003 } 1004 1005 CompileLog* log = C->log(); 1006 if (log != NULL) { 1007 log->head("eliminate_allocation type='%d'", 1008 log->identify(tklass->klass())); 1009 JVMState* p = alloc->jvms(); 1010 while (p != NULL) { 1011 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1012 p = p->caller(); 1013 } 1014 log->tail("eliminate_allocation"); 1015 } 1016 1017 process_users_of_allocation(alloc); 1018 1019 #ifndef PRODUCT 1020 if (PrintEliminateAllocations) { 1021 if (alloc->is_AllocateArray()) 1022 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1023 else 1024 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1025 } 1026 #endif 1027 1028 return true; 1029 } 1030 1031 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1032 // EA should remove all uses of non-escaping boxing node. 1033 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1034 return false; 1035 } 1036 1037 extract_call_projections(boxing); 1038 1039 const TypeTuple* r = boxing->tf()->range(); 1040 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1041 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1042 assert(t != NULL, "sanity"); 1043 1044 CompileLog* log = C->log(); 1045 if (log != NULL) { 1046 log->head("eliminate_boxing type='%d'", 1047 log->identify(t->klass())); 1048 JVMState* p = boxing->jvms(); 1049 while (p != NULL) { 1050 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1051 p = p->caller(); 1052 } 1053 log->tail("eliminate_boxing"); 1054 } 1055 1056 process_users_of_allocation(boxing); 1057 1058 #ifndef PRODUCT 1059 if (PrintEliminateAllocations) { 1060 tty->print("++++ Eliminated: %d ", boxing->_idx); 1061 boxing->method()->print_short_name(tty); 1062 tty->cr(); 1063 } 1064 #endif 1065 1066 return true; 1067 } 1068 1069 //---------------------------set_eden_pointers------------------------- 1070 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1071 if (UseTLAB) { // Private allocation: load from TLS 1072 Node* thread = transform_later(new ThreadLocalNode()); 1073 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1074 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1075 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1076 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1077 } else { // Shared allocation: load from globals 1078 CollectedHeap* ch = Universe::heap(); 1079 address top_adr = (address)ch->top_addr(); 1080 address end_adr = (address)ch->end_addr(); 1081 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1082 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1083 } 1084 } 1085 1086 1087 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1088 Node* adr = basic_plus_adr(base, offset); 1089 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1090 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1091 transform_later(value); 1092 return value; 1093 } 1094 1095 1096 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1097 Node* adr = basic_plus_adr(base, offset); 1098 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1099 transform_later(mem); 1100 return mem; 1101 } 1102 1103 //============================================================================= 1104 // 1105 // A L L O C A T I O N 1106 // 1107 // Allocation attempts to be fast in the case of frequent small objects. 1108 // It breaks down like this: 1109 // 1110 // 1) Size in doublewords is computed. This is a constant for objects and 1111 // variable for most arrays. Doubleword units are used to avoid size 1112 // overflow of huge doubleword arrays. We need doublewords in the end for 1113 // rounding. 1114 // 1115 // 2) Size is checked for being 'too large'. Too-large allocations will go 1116 // the slow path into the VM. The slow path can throw any required 1117 // exceptions, and does all the special checks for very large arrays. The 1118 // size test can constant-fold away for objects. For objects with 1119 // finalizers it constant-folds the otherway: you always go slow with 1120 // finalizers. 1121 // 1122 // 3) If NOT using TLABs, this is the contended loop-back point. 1123 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1124 // 1125 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1126 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1127 // "size*8" we always enter the VM, where "largish" is a constant picked small 1128 // enough that there's always space between the eden max and 4Gig (old space is 1129 // there so it's quite large) and large enough that the cost of entering the VM 1130 // is dwarfed by the cost to initialize the space. 1131 // 1132 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1133 // down. If contended, repeat at step 3. If using TLABs normal-store 1134 // adjusted heap top back down; there is no contention. 1135 // 1136 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1137 // fields. 1138 // 1139 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1140 // oop flavor. 1141 // 1142 //============================================================================= 1143 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1144 // Allocations bigger than this always go the slow route. 1145 // This value must be small enough that allocation attempts that need to 1146 // trigger exceptions go the slow route. Also, it must be small enough so 1147 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1148 //=============================================================================j// 1149 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1150 // The allocator will coalesce int->oop copies away. See comment in 1151 // coalesce.cpp about how this works. It depends critically on the exact 1152 // code shape produced here, so if you are changing this code shape 1153 // make sure the GC info for the heap-top is correct in and around the 1154 // slow-path call. 1155 // 1156 1157 void PhaseMacroExpand::expand_allocate_common( 1158 AllocateNode* alloc, // allocation node to be expanded 1159 Node* length, // array length for an array allocation 1160 const TypeFunc* slow_call_type, // Type of slow call 1161 address slow_call_address // Address of slow call 1162 ) 1163 { 1164 1165 Node* ctrl = alloc->in(TypeFunc::Control); 1166 Node* mem = alloc->in(TypeFunc::Memory); 1167 Node* i_o = alloc->in(TypeFunc::I_O); 1168 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1169 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1170 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1171 1172 assert(ctrl != NULL, "must have control"); 1173 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1174 // they will not be used if "always_slow" is set 1175 enum { slow_result_path = 1, fast_result_path = 2 }; 1176 Node *result_region; 1177 Node *result_phi_rawmem; 1178 Node *result_phi_rawoop; 1179 Node *result_phi_i_o; 1180 1181 // The initial slow comparison is a size check, the comparison 1182 // we want to do is a BoolTest::gt 1183 bool always_slow = false; 1184 int tv = _igvn.find_int_con(initial_slow_test, -1); 1185 if (tv >= 0) { 1186 always_slow = (tv == 1); 1187 initial_slow_test = NULL; 1188 } else { 1189 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1190 } 1191 1192 if (C->env()->dtrace_alloc_probes() || 1193 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() || 1194 (UseConcMarkSweepGC && CMSIncrementalMode))) { 1195 // Force slow-path allocation 1196 always_slow = true; 1197 initial_slow_test = NULL; 1198 } 1199 1200 1201 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1202 Node *slow_region = NULL; 1203 Node *toobig_false = ctrl; 1204 1205 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1206 // generate the initial test if necessary 1207 if (initial_slow_test != NULL ) { 1208 slow_region = new RegionNode(3); 1209 1210 // Now make the initial failure test. Usually a too-big test but 1211 // might be a TRUE for finalizers or a fancy class check for 1212 // newInstance0. 1213 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1214 transform_later(toobig_iff); 1215 // Plug the failing-too-big test into the slow-path region 1216 Node *toobig_true = new IfTrueNode( toobig_iff ); 1217 transform_later(toobig_true); 1218 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1219 toobig_false = new IfFalseNode( toobig_iff ); 1220 transform_later(toobig_false); 1221 } else { // No initial test, just fall into next case 1222 toobig_false = ctrl; 1223 debug_only(slow_region = NodeSentinel); 1224 } 1225 1226 Node *slow_mem = mem; // save the current memory state for slow path 1227 // generate the fast allocation code unless we know that the initial test will always go slow 1228 if (!always_slow) { 1229 // Fast path modifies only raw memory. 1230 if (mem->is_MergeMem()) { 1231 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1232 } 1233 1234 Node* eden_top_adr; 1235 Node* eden_end_adr; 1236 1237 set_eden_pointers(eden_top_adr, eden_end_adr); 1238 1239 // Load Eden::end. Loop invariant and hoisted. 1240 // 1241 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1242 // a TLAB to work around a bug where these values were being moved across 1243 // a safepoint. These are not oops, so they cannot be include in the oop 1244 // map, but they can be changed by a GC. The proper way to fix this would 1245 // be to set the raw memory state when generating a SafepointNode. However 1246 // this will require extensive changes to the loop optimization in order to 1247 // prevent a degradation of the optimization. 1248 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1249 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1250 1251 // allocate the Region and Phi nodes for the result 1252 result_region = new RegionNode(3); 1253 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1254 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1255 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1256 1257 // We need a Region for the loop-back contended case. 1258 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1259 Node *contended_region; 1260 Node *contended_phi_rawmem; 1261 if (UseTLAB) { 1262 contended_region = toobig_false; 1263 contended_phi_rawmem = mem; 1264 } else { 1265 contended_region = new RegionNode(3); 1266 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1267 // Now handle the passing-too-big test. We fall into the contended 1268 // loop-back merge point. 1269 contended_region ->init_req(fall_in_path, toobig_false); 1270 contended_phi_rawmem->init_req(fall_in_path, mem); 1271 transform_later(contended_region); 1272 transform_later(contended_phi_rawmem); 1273 } 1274 1275 // Load(-locked) the heap top. 1276 // See note above concerning the control input when using a TLAB 1277 Node *old_eden_top = UseTLAB 1278 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1279 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1280 1281 transform_later(old_eden_top); 1282 // Add to heap top to get a new heap top 1283 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1284 transform_later(new_eden_top); 1285 // Check for needing a GC; compare against heap end 1286 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1287 transform_later(needgc_cmp); 1288 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1289 transform_later(needgc_bol); 1290 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1291 transform_later(needgc_iff); 1292 1293 // Plug the failing-heap-space-need-gc test into the slow-path region 1294 Node *needgc_true = new IfTrueNode(needgc_iff); 1295 transform_later(needgc_true); 1296 if (initial_slow_test) { 1297 slow_region->init_req(need_gc_path, needgc_true); 1298 // This completes all paths into the slow merge point 1299 transform_later(slow_region); 1300 } else { // No initial slow path needed! 1301 // Just fall from the need-GC path straight into the VM call. 1302 slow_region = needgc_true; 1303 } 1304 // No need for a GC. Setup for the Store-Conditional 1305 Node *needgc_false = new IfFalseNode(needgc_iff); 1306 transform_later(needgc_false); 1307 1308 // Grab regular I/O before optional prefetch may change it. 1309 // Slow-path does no I/O so just set it to the original I/O. 1310 result_phi_i_o->init_req(slow_result_path, i_o); 1311 1312 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1313 old_eden_top, new_eden_top, length); 1314 1315 // Name successful fast-path variables 1316 Node* fast_oop = old_eden_top; 1317 Node* fast_oop_ctrl; 1318 Node* fast_oop_rawmem; 1319 1320 // Store (-conditional) the modified eden top back down. 1321 // StorePConditional produces flags for a test PLUS a modified raw 1322 // memory state. 1323 if (UseTLAB) { 1324 Node* store_eden_top = 1325 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1326 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1327 transform_later(store_eden_top); 1328 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1329 fast_oop_rawmem = store_eden_top; 1330 } else { 1331 Node* store_eden_top = 1332 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1333 new_eden_top, fast_oop/*old_eden_top*/); 1334 transform_later(store_eden_top); 1335 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1336 transform_later(contention_check); 1337 store_eden_top = new SCMemProjNode(store_eden_top); 1338 transform_later(store_eden_top); 1339 1340 // If not using TLABs, check to see if there was contention. 1341 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1342 transform_later(contention_iff); 1343 Node *contention_true = new IfTrueNode(contention_iff); 1344 transform_later(contention_true); 1345 // If contention, loopback and try again. 1346 contended_region->init_req(contended_loopback_path, contention_true); 1347 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1348 1349 // Fast-path succeeded with no contention! 1350 Node *contention_false = new IfFalseNode(contention_iff); 1351 transform_later(contention_false); 1352 fast_oop_ctrl = contention_false; 1353 1354 // Bump total allocated bytes for this thread 1355 Node* thread = new ThreadLocalNode(); 1356 transform_later(thread); 1357 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1358 in_bytes(JavaThread::allocated_bytes_offset())); 1359 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1360 0, TypeLong::LONG, T_LONG); 1361 #ifdef _LP64 1362 Node* alloc_size = size_in_bytes; 1363 #else 1364 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1365 transform_later(alloc_size); 1366 #endif 1367 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1368 transform_later(new_alloc_bytes); 1369 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1370 0, new_alloc_bytes, T_LONG); 1371 } 1372 1373 InitializeNode* init = alloc->initialization(); 1374 fast_oop_rawmem = initialize_object(alloc, 1375 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1376 klass_node, length, size_in_bytes); 1377 1378 // If initialization is performed by an array copy, any required 1379 // MemBarStoreStore was already added. If the object does not 1380 // escape no need for a MemBarStoreStore. Otherwise we need a 1381 // MemBarStoreStore so that stores that initialize this object 1382 // can't be reordered with a subsequent store that makes this 1383 // object accessible by other threads. 1384 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) { 1385 if (init == NULL || init->req() < InitializeNode::RawStores) { 1386 // No InitializeNode or no stores captured by zeroing 1387 // elimination. Simply add the MemBarStoreStore after object 1388 // initialization. 1389 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1390 transform_later(mb); 1391 1392 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1393 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1394 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1395 transform_later(fast_oop_ctrl); 1396 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1397 transform_later(fast_oop_rawmem); 1398 } else { 1399 // Add the MemBarStoreStore after the InitializeNode so that 1400 // all stores performing the initialization that were moved 1401 // before the InitializeNode happen before the storestore 1402 // barrier. 1403 1404 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1405 Node* init_mem = init->proj_out(TypeFunc::Memory); 1406 1407 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1408 transform_later(mb); 1409 1410 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1411 transform_later(ctrl); 1412 Node* mem = new ProjNode(init,TypeFunc::Memory); 1413 transform_later(mem); 1414 1415 // The MemBarStoreStore depends on control and memory coming 1416 // from the InitializeNode 1417 mb->init_req(TypeFunc::Memory, mem); 1418 mb->init_req(TypeFunc::Control, ctrl); 1419 1420 ctrl = new ProjNode(mb,TypeFunc::Control); 1421 transform_later(ctrl); 1422 mem = new ProjNode(mb,TypeFunc::Memory); 1423 transform_later(mem); 1424 1425 // All nodes that depended on the InitializeNode for control 1426 // and memory must now depend on the MemBarNode that itself 1427 // depends on the InitializeNode 1428 _igvn.replace_node(init_ctrl, ctrl); 1429 _igvn.replace_node(init_mem, mem); 1430 } 1431 } 1432 1433 if (C->env()->dtrace_extended_probes()) { 1434 // Slow-path call 1435 int size = TypeFunc::Parms + 2; 1436 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1437 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1438 "dtrace_object_alloc", 1439 TypeRawPtr::BOTTOM); 1440 1441 // Get base of thread-local storage area 1442 Node* thread = new ThreadLocalNode(); 1443 transform_later(thread); 1444 1445 call->init_req(TypeFunc::Parms+0, thread); 1446 call->init_req(TypeFunc::Parms+1, fast_oop); 1447 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1448 call->init_req(TypeFunc::I_O , top()); // does no i/o 1449 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1450 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1451 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1452 transform_later(call); 1453 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1454 transform_later(fast_oop_ctrl); 1455 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1456 transform_later(fast_oop_rawmem); 1457 } 1458 1459 // Plug in the successful fast-path into the result merge point 1460 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1461 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1462 result_phi_i_o ->init_req(fast_result_path, i_o); 1463 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1464 } else { 1465 slow_region = ctrl; 1466 result_phi_i_o = i_o; // Rename it to use in the following code. 1467 } 1468 1469 // Generate slow-path call 1470 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1471 OptoRuntime::stub_name(slow_call_address), 1472 alloc->jvms()->bci(), 1473 TypePtr::BOTTOM); 1474 call->init_req( TypeFunc::Control, slow_region ); 1475 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1476 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1477 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1478 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1479 1480 call->init_req(TypeFunc::Parms+0, klass_node); 1481 if (length != NULL) { 1482 call->init_req(TypeFunc::Parms+1, length); 1483 } 1484 1485 // Copy debug information and adjust JVMState information, then replace 1486 // allocate node with the call 1487 copy_call_debug_info((CallNode *) alloc, call); 1488 if (!always_slow) { 1489 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1490 } else { 1491 // Hook i_o projection to avoid its elimination during allocation 1492 // replacement (when only a slow call is generated). 1493 call->set_req(TypeFunc::I_O, result_phi_i_o); 1494 } 1495 _igvn.replace_node(alloc, call); 1496 transform_later(call); 1497 1498 // Identify the output projections from the allocate node and 1499 // adjust any references to them. 1500 // The control and io projections look like: 1501 // 1502 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1503 // Allocate Catch 1504 // ^---Proj(io) <-------+ ^---CatchProj(io) 1505 // 1506 // We are interested in the CatchProj nodes. 1507 // 1508 extract_call_projections(call); 1509 1510 // An allocate node has separate memory projections for the uses on 1511 // the control and i_o paths. Replace the control memory projection with 1512 // result_phi_rawmem (unless we are only generating a slow call when 1513 // both memory projections are combined) 1514 if (!always_slow && _memproj_fallthrough != NULL) { 1515 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1516 Node *use = _memproj_fallthrough->fast_out(i); 1517 _igvn.rehash_node_delayed(use); 1518 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1519 // back up iterator 1520 --i; 1521 } 1522 } 1523 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1524 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1525 if (_memproj_catchall != NULL ) { 1526 if (_memproj_fallthrough == NULL) { 1527 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1528 transform_later(_memproj_fallthrough); 1529 } 1530 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1531 Node *use = _memproj_catchall->fast_out(i); 1532 _igvn.rehash_node_delayed(use); 1533 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1534 // back up iterator 1535 --i; 1536 } 1537 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1538 _igvn.remove_dead_node(_memproj_catchall); 1539 } 1540 1541 // An allocate node has separate i_o projections for the uses on the control 1542 // and i_o paths. Always replace the control i_o projection with result i_o 1543 // otherwise incoming i_o become dead when only a slow call is generated 1544 // (it is different from memory projections where both projections are 1545 // combined in such case). 1546 if (_ioproj_fallthrough != NULL) { 1547 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1548 Node *use = _ioproj_fallthrough->fast_out(i); 1549 _igvn.rehash_node_delayed(use); 1550 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1551 // back up iterator 1552 --i; 1553 } 1554 } 1555 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1556 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1557 if (_ioproj_catchall != NULL ) { 1558 if (_ioproj_fallthrough == NULL) { 1559 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1560 transform_later(_ioproj_fallthrough); 1561 } 1562 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1563 Node *use = _ioproj_catchall->fast_out(i); 1564 _igvn.rehash_node_delayed(use); 1565 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1566 // back up iterator 1567 --i; 1568 } 1569 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1570 _igvn.remove_dead_node(_ioproj_catchall); 1571 } 1572 1573 // if we generated only a slow call, we are done 1574 if (always_slow) { 1575 // Now we can unhook i_o. 1576 if (result_phi_i_o->outcnt() > 1) { 1577 call->set_req(TypeFunc::I_O, top()); 1578 } else { 1579 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1580 // Case of new array with negative size known during compilation. 1581 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1582 // following code since call to runtime will throw exception. 1583 // As result there will be no users of i_o after the call. 1584 // Leave i_o attached to this call to avoid problems in preceding graph. 1585 } 1586 return; 1587 } 1588 1589 1590 if (_fallthroughcatchproj != NULL) { 1591 ctrl = _fallthroughcatchproj->clone(); 1592 transform_later(ctrl); 1593 _igvn.replace_node(_fallthroughcatchproj, result_region); 1594 } else { 1595 ctrl = top(); 1596 } 1597 Node *slow_result; 1598 if (_resproj == NULL) { 1599 // no uses of the allocation result 1600 slow_result = top(); 1601 } else { 1602 slow_result = _resproj->clone(); 1603 transform_later(slow_result); 1604 _igvn.replace_node(_resproj, result_phi_rawoop); 1605 } 1606 1607 // Plug slow-path into result merge point 1608 result_region ->init_req( slow_result_path, ctrl ); 1609 result_phi_rawoop->init_req( slow_result_path, slow_result); 1610 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1611 transform_later(result_region); 1612 transform_later(result_phi_rawoop); 1613 transform_later(result_phi_rawmem); 1614 transform_later(result_phi_i_o); 1615 // This completes all paths into the result merge point 1616 } 1617 1618 1619 // Helper for PhaseMacroExpand::expand_allocate_common. 1620 // Initializes the newly-allocated storage. 1621 Node* 1622 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1623 Node* control, Node* rawmem, Node* object, 1624 Node* klass_node, Node* length, 1625 Node* size_in_bytes) { 1626 InitializeNode* init = alloc->initialization(); 1627 // Store the klass & mark bits 1628 Node* mark_node = NULL; 1629 // For now only enable fast locking for non-array types 1630 if (UseBiasedLocking && (length == NULL)) { 1631 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1632 } else { 1633 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1634 } 1635 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1636 1637 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1638 int header_size = alloc->minimum_header_size(); // conservatively small 1639 1640 // Array length 1641 if (length != NULL) { // Arrays need length field 1642 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1643 // conservatively small header size: 1644 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1645 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1646 if (k->is_array_klass()) // we know the exact header size in most cases: 1647 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1648 } 1649 1650 // Clear the object body, if necessary. 1651 if (init == NULL) { 1652 // The init has somehow disappeared; be cautious and clear everything. 1653 // 1654 // This can happen if a node is allocated but an uncommon trap occurs 1655 // immediately. In this case, the Initialize gets associated with the 1656 // trap, and may be placed in a different (outer) loop, if the Allocate 1657 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1658 // there can be two Allocates to one Initialize. The answer in all these 1659 // edge cases is safety first. It is always safe to clear immediately 1660 // within an Allocate, and then (maybe or maybe not) clear some more later. 1661 if (!ZeroTLAB) 1662 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1663 header_size, size_in_bytes, 1664 &_igvn); 1665 } else { 1666 if (!init->is_complete()) { 1667 // Try to win by zeroing only what the init does not store. 1668 // We can also try to do some peephole optimizations, 1669 // such as combining some adjacent subword stores. 1670 rawmem = init->complete_stores(control, rawmem, object, 1671 header_size, size_in_bytes, &_igvn); 1672 } 1673 // We have no more use for this link, since the AllocateNode goes away: 1674 init->set_req(InitializeNode::RawAddress, top()); 1675 // (If we keep the link, it just confuses the register allocator, 1676 // who thinks he sees a real use of the address by the membar.) 1677 } 1678 1679 return rawmem; 1680 } 1681 1682 // Generate prefetch instructions for next allocations. 1683 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1684 Node*& contended_phi_rawmem, 1685 Node* old_eden_top, Node* new_eden_top, 1686 Node* length) { 1687 enum { fall_in_path = 1, pf_path = 2 }; 1688 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1689 // Generate prefetch allocation with watermark check. 1690 // As an allocation hits the watermark, we will prefetch starting 1691 // at a "distance" away from watermark. 1692 1693 Node *pf_region = new RegionNode(3); 1694 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1695 TypeRawPtr::BOTTOM ); 1696 // I/O is used for Prefetch 1697 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1698 1699 Node *thread = new ThreadLocalNode(); 1700 transform_later(thread); 1701 1702 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1703 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1704 transform_later(eden_pf_adr); 1705 1706 Node *old_pf_wm = new LoadPNode(needgc_false, 1707 contended_phi_rawmem, eden_pf_adr, 1708 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1709 MemNode::unordered); 1710 transform_later(old_pf_wm); 1711 1712 // check against new_eden_top 1713 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1714 transform_later(need_pf_cmp); 1715 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1716 transform_later(need_pf_bol); 1717 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1718 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1719 transform_later(need_pf_iff); 1720 1721 // true node, add prefetchdistance 1722 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1723 transform_later(need_pf_true); 1724 1725 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1726 transform_later(need_pf_false); 1727 1728 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1729 _igvn.MakeConX(AllocatePrefetchDistance) ); 1730 transform_later(new_pf_wmt ); 1731 new_pf_wmt->set_req(0, need_pf_true); 1732 1733 Node *store_new_wmt = new StorePNode(need_pf_true, 1734 contended_phi_rawmem, eden_pf_adr, 1735 TypeRawPtr::BOTTOM, new_pf_wmt, 1736 MemNode::unordered); 1737 transform_later(store_new_wmt); 1738 1739 // adding prefetches 1740 pf_phi_abio->init_req( fall_in_path, i_o ); 1741 1742 Node *prefetch_adr; 1743 Node *prefetch; 1744 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1745 uint step_size = AllocatePrefetchStepSize; 1746 uint distance = 0; 1747 1748 for ( uint i = 0; i < lines; i++ ) { 1749 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1750 _igvn.MakeConX(distance) ); 1751 transform_later(prefetch_adr); 1752 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1753 transform_later(prefetch); 1754 distance += step_size; 1755 i_o = prefetch; 1756 } 1757 pf_phi_abio->set_req( pf_path, i_o ); 1758 1759 pf_region->init_req( fall_in_path, need_pf_false ); 1760 pf_region->init_req( pf_path, need_pf_true ); 1761 1762 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1763 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1764 1765 transform_later(pf_region); 1766 transform_later(pf_phi_rawmem); 1767 transform_later(pf_phi_abio); 1768 1769 needgc_false = pf_region; 1770 contended_phi_rawmem = pf_phi_rawmem; 1771 i_o = pf_phi_abio; 1772 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1773 // Insert a prefetch for each allocation. 1774 // This code is used for Sparc with BIS. 1775 Node *pf_region = new RegionNode(3); 1776 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1777 TypeRawPtr::BOTTOM ); 1778 1779 // Generate several prefetch instructions. 1780 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1781 uint step_size = AllocatePrefetchStepSize; 1782 uint distance = AllocatePrefetchDistance; 1783 1784 // Next cache address. 1785 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1786 _igvn.MakeConX(distance)); 1787 transform_later(cache_adr); 1788 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1789 transform_later(cache_adr); 1790 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1791 cache_adr = new AndXNode(cache_adr, mask); 1792 transform_later(cache_adr); 1793 cache_adr = new CastX2PNode(cache_adr); 1794 transform_later(cache_adr); 1795 1796 // Prefetch 1797 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1798 prefetch->set_req(0, needgc_false); 1799 transform_later(prefetch); 1800 contended_phi_rawmem = prefetch; 1801 Node *prefetch_adr; 1802 distance = step_size; 1803 for ( uint i = 1; i < lines; i++ ) { 1804 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1805 _igvn.MakeConX(distance) ); 1806 transform_later(prefetch_adr); 1807 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1808 transform_later(prefetch); 1809 distance += step_size; 1810 contended_phi_rawmem = prefetch; 1811 } 1812 } else if( AllocatePrefetchStyle > 0 ) { 1813 // Insert a prefetch for each allocation only on the fast-path 1814 Node *prefetch_adr; 1815 Node *prefetch; 1816 // Generate several prefetch instructions. 1817 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1818 uint step_size = AllocatePrefetchStepSize; 1819 uint distance = AllocatePrefetchDistance; 1820 for ( uint i = 0; i < lines; i++ ) { 1821 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1822 _igvn.MakeConX(distance) ); 1823 transform_later(prefetch_adr); 1824 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1825 // Do not let it float too high, since if eden_top == eden_end, 1826 // both might be null. 1827 if( i == 0 ) { // Set control for first prefetch, next follows it 1828 prefetch->init_req(0, needgc_false); 1829 } 1830 transform_later(prefetch); 1831 distance += step_size; 1832 i_o = prefetch; 1833 } 1834 } 1835 return i_o; 1836 } 1837 1838 1839 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1840 expand_allocate_common(alloc, NULL, 1841 OptoRuntime::new_instance_Type(), 1842 OptoRuntime::new_instance_Java()); 1843 } 1844 1845 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1846 Node* length = alloc->in(AllocateNode::ALength); 1847 InitializeNode* init = alloc->initialization(); 1848 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1849 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1850 address slow_call_address; // Address of slow call 1851 if (init != NULL && init->is_complete_with_arraycopy() && 1852 k->is_type_array_klass()) { 1853 // Don't zero type array during slow allocation in VM since 1854 // it will be initialized later by arraycopy in compiled code. 1855 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1856 } else { 1857 slow_call_address = OptoRuntime::new_array_Java(); 1858 } 1859 expand_allocate_common(alloc, length, 1860 OptoRuntime::new_array_Type(), 1861 slow_call_address); 1862 } 1863 1864 //-------------------mark_eliminated_box---------------------------------- 1865 // 1866 // During EA obj may point to several objects but after few ideal graph 1867 // transformations (CCP) it may point to only one non escaping object 1868 // (but still using phi), corresponding locks and unlocks will be marked 1869 // for elimination. Later obj could be replaced with a new node (new phi) 1870 // and which does not have escape information. And later after some graph 1871 // reshape other locks and unlocks (which were not marked for elimination 1872 // before) are connected to this new obj (phi) but they still will not be 1873 // marked for elimination since new obj has no escape information. 1874 // Mark all associated (same box and obj) lock and unlock nodes for 1875 // elimination if some of them marked already. 1876 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1877 if (oldbox->as_BoxLock()->is_eliminated()) 1878 return; // This BoxLock node was processed already. 1879 1880 // New implementation (EliminateNestedLocks) has separate BoxLock 1881 // node for each locked region so mark all associated locks/unlocks as 1882 // eliminated even if different objects are referenced in one locked region 1883 // (for example, OSR compilation of nested loop inside locked scope). 1884 if (EliminateNestedLocks || 1885 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 1886 // Box is used only in one lock region. Mark this box as eliminated. 1887 _igvn.hash_delete(oldbox); 1888 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1889 _igvn.hash_insert(oldbox); 1890 1891 for (uint i = 0; i < oldbox->outcnt(); i++) { 1892 Node* u = oldbox->raw_out(i); 1893 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1894 AbstractLockNode* alock = u->as_AbstractLock(); 1895 // Check lock's box since box could be referenced by Lock's debug info. 1896 if (alock->box_node() == oldbox) { 1897 // Mark eliminated all related locks and unlocks. 1898 alock->set_non_esc_obj(); 1899 } 1900 } 1901 } 1902 return; 1903 } 1904 1905 // Create new "eliminated" BoxLock node and use it in monitor debug info 1906 // instead of oldbox for the same object. 1907 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1908 1909 // Note: BoxLock node is marked eliminated only here and it is used 1910 // to indicate that all associated lock and unlock nodes are marked 1911 // for elimination. 1912 newbox->set_eliminated(); 1913 transform_later(newbox); 1914 1915 // Replace old box node with new box for all users of the same object. 1916 for (uint i = 0; i < oldbox->outcnt();) { 1917 bool next_edge = true; 1918 1919 Node* u = oldbox->raw_out(i); 1920 if (u->is_AbstractLock()) { 1921 AbstractLockNode* alock = u->as_AbstractLock(); 1922 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 1923 // Replace Box and mark eliminated all related locks and unlocks. 1924 alock->set_non_esc_obj(); 1925 _igvn.rehash_node_delayed(alock); 1926 alock->set_box_node(newbox); 1927 next_edge = false; 1928 } 1929 } 1930 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 1931 FastLockNode* flock = u->as_FastLock(); 1932 assert(flock->box_node() == oldbox, "sanity"); 1933 _igvn.rehash_node_delayed(flock); 1934 flock->set_box_node(newbox); 1935 next_edge = false; 1936 } 1937 1938 // Replace old box in monitor debug info. 1939 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1940 SafePointNode* sfn = u->as_SafePoint(); 1941 JVMState* youngest_jvms = sfn->jvms(); 1942 int max_depth = youngest_jvms->depth(); 1943 for (int depth = 1; depth <= max_depth; depth++) { 1944 JVMState* jvms = youngest_jvms->of_depth(depth); 1945 int num_mon = jvms->nof_monitors(); 1946 // Loop over monitors 1947 for (int idx = 0; idx < num_mon; idx++) { 1948 Node* obj_node = sfn->monitor_obj(jvms, idx); 1949 Node* box_node = sfn->monitor_box(jvms, idx); 1950 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 1951 int j = jvms->monitor_box_offset(idx); 1952 _igvn.replace_input_of(u, j, newbox); 1953 next_edge = false; 1954 } 1955 } 1956 } 1957 } 1958 if (next_edge) i++; 1959 } 1960 } 1961 1962 //-----------------------mark_eliminated_locking_nodes----------------------- 1963 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 1964 if (EliminateNestedLocks) { 1965 if (alock->is_nested()) { 1966 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 1967 return; 1968 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 1969 // Only Lock node has JVMState needed here. 1970 if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) { 1971 // Mark eliminated related nested locks and unlocks. 1972 Node* obj = alock->obj_node(); 1973 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 1974 assert(!box_node->is_eliminated(), "should not be marked yet"); 1975 // Note: BoxLock node is marked eliminated only here 1976 // and it is used to indicate that all associated lock 1977 // and unlock nodes are marked for elimination. 1978 box_node->set_eliminated(); // Box's hash is always NO_HASH here 1979 for (uint i = 0; i < box_node->outcnt(); i++) { 1980 Node* u = box_node->raw_out(i); 1981 if (u->is_AbstractLock()) { 1982 alock = u->as_AbstractLock(); 1983 if (alock->box_node() == box_node) { 1984 // Verify that this Box is referenced only by related locks. 1985 assert(alock->obj_node()->eqv_uncast(obj), ""); 1986 // Mark all related locks and unlocks. 1987 alock->set_nested(); 1988 } 1989 } 1990 } 1991 } 1992 return; 1993 } 1994 // Process locks for non escaping object 1995 assert(alock->is_non_esc_obj(), ""); 1996 } // EliminateNestedLocks 1997 1998 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 1999 // Look for all locks of this object and mark them and 2000 // corresponding BoxLock nodes as eliminated. 2001 Node* obj = alock->obj_node(); 2002 for (uint j = 0; j < obj->outcnt(); j++) { 2003 Node* o = obj->raw_out(j); 2004 if (o->is_AbstractLock() && 2005 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2006 alock = o->as_AbstractLock(); 2007 Node* box = alock->box_node(); 2008 // Replace old box node with new eliminated box for all users 2009 // of the same object and mark related locks as eliminated. 2010 mark_eliminated_box(box, obj); 2011 } 2012 } 2013 } 2014 } 2015 2016 // we have determined that this lock/unlock can be eliminated, we simply 2017 // eliminate the node without expanding it. 2018 // 2019 // Note: The membar's associated with the lock/unlock are currently not 2020 // eliminated. This should be investigated as a future enhancement. 2021 // 2022 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2023 2024 if (!alock->is_eliminated()) { 2025 return false; 2026 } 2027 #ifdef ASSERT 2028 if (!alock->is_coarsened()) { 2029 // Check that new "eliminated" BoxLock node is created. 2030 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2031 assert(oldbox->is_eliminated(), "should be done already"); 2032 } 2033 #endif 2034 CompileLog* log = C->log(); 2035 if (log != NULL) { 2036 log->head("eliminate_lock lock='%d'", 2037 alock->is_Lock()); 2038 JVMState* p = alock->jvms(); 2039 while (p != NULL) { 2040 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 2041 p = p->caller(); 2042 } 2043 log->tail("eliminate_lock"); 2044 } 2045 2046 #ifndef PRODUCT 2047 if (PrintEliminateLocks) { 2048 if (alock->is_Lock()) { 2049 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2050 } else { 2051 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2052 } 2053 } 2054 #endif 2055 2056 Node* mem = alock->in(TypeFunc::Memory); 2057 Node* ctrl = alock->in(TypeFunc::Control); 2058 2059 extract_call_projections(alock); 2060 // There are 2 projections from the lock. The lock node will 2061 // be deleted when its last use is subsumed below. 2062 assert(alock->outcnt() == 2 && 2063 _fallthroughproj != NULL && 2064 _memproj_fallthrough != NULL, 2065 "Unexpected projections from Lock/Unlock"); 2066 2067 Node* fallthroughproj = _fallthroughproj; 2068 Node* memproj_fallthrough = _memproj_fallthrough; 2069 2070 // The memory projection from a lock/unlock is RawMem 2071 // The input to a Lock is merged memory, so extract its RawMem input 2072 // (unless the MergeMem has been optimized away.) 2073 if (alock->is_Lock()) { 2074 // Seach for MemBarAcquireLock node and delete it also. 2075 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2076 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2077 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2078 Node* memproj = membar->proj_out(TypeFunc::Memory); 2079 _igvn.replace_node(ctrlproj, fallthroughproj); 2080 _igvn.replace_node(memproj, memproj_fallthrough); 2081 2082 // Delete FastLock node also if this Lock node is unique user 2083 // (a loop peeling may clone a Lock node). 2084 Node* flock = alock->as_Lock()->fastlock_node(); 2085 if (flock->outcnt() == 1) { 2086 assert(flock->unique_out() == alock, "sanity"); 2087 _igvn.replace_node(flock, top()); 2088 } 2089 } 2090 2091 // Seach for MemBarReleaseLock node and delete it also. 2092 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2093 ctrl->in(0)->is_MemBar()) { 2094 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2095 assert(membar->Opcode() == Op_MemBarReleaseLock && 2096 mem->is_Proj() && membar == mem->in(0), ""); 2097 _igvn.replace_node(fallthroughproj, ctrl); 2098 _igvn.replace_node(memproj_fallthrough, mem); 2099 fallthroughproj = ctrl; 2100 memproj_fallthrough = mem; 2101 ctrl = membar->in(TypeFunc::Control); 2102 mem = membar->in(TypeFunc::Memory); 2103 } 2104 2105 _igvn.replace_node(fallthroughproj, ctrl); 2106 _igvn.replace_node(memproj_fallthrough, mem); 2107 return true; 2108 } 2109 2110 2111 //------------------------------expand_lock_node---------------------- 2112 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2113 2114 Node* ctrl = lock->in(TypeFunc::Control); 2115 Node* mem = lock->in(TypeFunc::Memory); 2116 Node* obj = lock->obj_node(); 2117 Node* box = lock->box_node(); 2118 Node* flock = lock->fastlock_node(); 2119 2120 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2121 2122 // Make the merge point 2123 Node *region; 2124 Node *mem_phi; 2125 Node *slow_path; 2126 2127 if (UseOptoBiasInlining) { 2128 /* 2129 * See the full description in MacroAssembler::biased_locking_enter(). 2130 * 2131 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2132 * // The object is biased. 2133 * proto_node = klass->prototype_header; 2134 * o_node = thread | proto_node; 2135 * x_node = o_node ^ mark_word; 2136 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2137 * // Done. 2138 * } else { 2139 * if( (x_node & biased_lock_mask) != 0 ) { 2140 * // The klass's prototype header is no longer biased. 2141 * cas(&mark_word, mark_word, proto_node) 2142 * goto cas_lock; 2143 * } else { 2144 * // The klass's prototype header is still biased. 2145 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2146 * old = mark_word; 2147 * new = o_node; 2148 * } else { 2149 * // Different thread or anonymous biased. 2150 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2151 * new = thread | old; 2152 * } 2153 * // Try to rebias. 2154 * if( cas(&mark_word, old, new) == 0 ) { 2155 * // Done. 2156 * } else { 2157 * goto slow_path; // Failed. 2158 * } 2159 * } 2160 * } 2161 * } else { 2162 * // The object is not biased. 2163 * cas_lock: 2164 * if( FastLock(obj) == 0 ) { 2165 * // Done. 2166 * } else { 2167 * slow_path: 2168 * OptoRuntime::complete_monitor_locking_Java(obj); 2169 * } 2170 * } 2171 */ 2172 2173 region = new RegionNode(5); 2174 // create a Phi for the memory state 2175 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2176 2177 Node* fast_lock_region = new RegionNode(3); 2178 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2179 2180 // First, check mark word for the biased lock pattern. 2181 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2182 2183 // Get fast path - mark word has the biased lock pattern. 2184 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2185 markOopDesc::biased_lock_mask_in_place, 2186 markOopDesc::biased_lock_pattern, true); 2187 // fast_lock_region->in(1) is set to slow path. 2188 fast_lock_mem_phi->init_req(1, mem); 2189 2190 // Now check that the lock is biased to the current thread and has 2191 // the same epoch and bias as Klass::_prototype_header. 2192 2193 // Special-case a fresh allocation to avoid building nodes: 2194 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2195 if (klass_node == NULL) { 2196 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2197 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) ); 2198 #ifdef _LP64 2199 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2200 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2201 klass_node->in(1)->init_req(0, ctrl); 2202 } else 2203 #endif 2204 klass_node->init_req(0, ctrl); 2205 } 2206 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2207 2208 Node* thread = transform_later(new ThreadLocalNode()); 2209 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2210 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2211 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2212 2213 // Get slow path - mark word does NOT match the value. 2214 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2215 (~markOopDesc::age_mask_in_place), 0); 2216 // region->in(3) is set to fast path - the object is biased to the current thread. 2217 mem_phi->init_req(3, mem); 2218 2219 2220 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2221 2222 2223 // First, check biased pattern. 2224 // Get fast path - _prototype_header has the same biased lock pattern. 2225 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2226 markOopDesc::biased_lock_mask_in_place, 0, true); 2227 2228 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2229 // fast_lock_region->in(2) - the prototype header is no longer biased 2230 // and we have to revoke the bias on this object. 2231 // We are going to try to reset the mark of this object to the prototype 2232 // value and fall through to the CAS-based locking scheme. 2233 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2234 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2235 proto_node, mark_node); 2236 transform_later(cas); 2237 Node* proj = transform_later(new SCMemProjNode(cas)); 2238 fast_lock_mem_phi->init_req(2, proj); 2239 2240 2241 // Second, check epoch bits. 2242 Node* rebiased_region = new RegionNode(3); 2243 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2244 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2245 2246 // Get slow path - mark word does NOT match epoch bits. 2247 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2248 markOopDesc::epoch_mask_in_place, 0); 2249 // The epoch of the current bias is not valid, attempt to rebias the object 2250 // toward the current thread. 2251 rebiased_region->init_req(2, epoch_ctrl); 2252 old_phi->init_req(2, mark_node); 2253 new_phi->init_req(2, o_node); 2254 2255 // rebiased_region->in(1) is set to fast path. 2256 // The epoch of the current bias is still valid but we know 2257 // nothing about the owner; it might be set or it might be clear. 2258 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2259 markOopDesc::age_mask_in_place | 2260 markOopDesc::epoch_mask_in_place); 2261 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2262 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2263 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2264 old_phi->init_req(1, old); 2265 new_phi->init_req(1, new_mark); 2266 2267 transform_later(rebiased_region); 2268 transform_later(old_phi); 2269 transform_later(new_phi); 2270 2271 // Try to acquire the bias of the object using an atomic operation. 2272 // If this fails we will go in to the runtime to revoke the object's bias. 2273 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2274 transform_later(cas); 2275 proj = transform_later(new SCMemProjNode(cas)); 2276 2277 // Get slow path - Failed to CAS. 2278 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2279 mem_phi->init_req(4, proj); 2280 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2281 2282 // Failed to CAS. 2283 slow_path = new RegionNode(3); 2284 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2285 2286 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2287 slow_mem->init_req(1, proj); 2288 2289 // Call CAS-based locking scheme (FastLock node). 2290 2291 transform_later(fast_lock_region); 2292 transform_later(fast_lock_mem_phi); 2293 2294 // Get slow path - FastLock failed to lock the object. 2295 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2296 mem_phi->init_req(2, fast_lock_mem_phi); 2297 // region->in(2) is set to fast path - the object is locked to the current thread. 2298 2299 slow_path->init_req(2, ctrl); // Capture slow-control 2300 slow_mem->init_req(2, fast_lock_mem_phi); 2301 2302 transform_later(slow_path); 2303 transform_later(slow_mem); 2304 // Reset lock's memory edge. 2305 lock->set_req(TypeFunc::Memory, slow_mem); 2306 2307 } else { 2308 region = new RegionNode(3); 2309 // create a Phi for the memory state 2310 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2311 2312 // Optimize test; set region slot 2 2313 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2314 mem_phi->init_req(2, mem); 2315 } 2316 2317 // Make slow path call 2318 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); 2319 2320 extract_call_projections(call); 2321 2322 // Slow path can only throw asynchronous exceptions, which are always 2323 // de-opted. So the compiler thinks the slow-call can never throw an 2324 // exception. If it DOES throw an exception we would need the debug 2325 // info removed first (since if it throws there is no monitor). 2326 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2327 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2328 2329 // Capture slow path 2330 // disconnect fall-through projection from call and create a new one 2331 // hook up users of fall-through projection to region 2332 Node *slow_ctrl = _fallthroughproj->clone(); 2333 transform_later(slow_ctrl); 2334 _igvn.hash_delete(_fallthroughproj); 2335 _fallthroughproj->disconnect_inputs(NULL, C); 2336 region->init_req(1, slow_ctrl); 2337 // region inputs are now complete 2338 transform_later(region); 2339 _igvn.replace_node(_fallthroughproj, region); 2340 2341 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2342 mem_phi->init_req(1, memproj ); 2343 transform_later(mem_phi); 2344 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2345 } 2346 2347 //------------------------------expand_unlock_node---------------------- 2348 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2349 2350 Node* ctrl = unlock->in(TypeFunc::Control); 2351 Node* mem = unlock->in(TypeFunc::Memory); 2352 Node* obj = unlock->obj_node(); 2353 Node* box = unlock->box_node(); 2354 2355 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2356 2357 // No need for a null check on unlock 2358 2359 // Make the merge point 2360 Node *region; 2361 Node *mem_phi; 2362 2363 if (UseOptoBiasInlining) { 2364 // Check for biased locking unlock case, which is a no-op. 2365 // See the full description in MacroAssembler::biased_locking_exit(). 2366 region = new RegionNode(4); 2367 // create a Phi for the memory state 2368 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2369 mem_phi->init_req(3, mem); 2370 2371 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2372 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2373 markOopDesc::biased_lock_mask_in_place, 2374 markOopDesc::biased_lock_pattern); 2375 } else { 2376 region = new RegionNode(3); 2377 // create a Phi for the memory state 2378 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2379 } 2380 2381 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2382 funlock = transform_later( funlock )->as_FastUnlock(); 2383 // Optimize test; set region slot 2 2384 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2385 2386 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box ); 2387 2388 extract_call_projections(call); 2389 2390 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2391 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2392 2393 // No exceptions for unlocking 2394 // Capture slow path 2395 // disconnect fall-through projection from call and create a new one 2396 // hook up users of fall-through projection to region 2397 Node *slow_ctrl = _fallthroughproj->clone(); 2398 transform_later(slow_ctrl); 2399 _igvn.hash_delete(_fallthroughproj); 2400 _fallthroughproj->disconnect_inputs(NULL, C); 2401 region->init_req(1, slow_ctrl); 2402 // region inputs are now complete 2403 transform_later(region); 2404 _igvn.replace_node(_fallthroughproj, region); 2405 2406 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2407 mem_phi->init_req(1, memproj ); 2408 mem_phi->init_req(2, mem); 2409 transform_later(mem_phi); 2410 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2411 } 2412 2413 //---------------------------eliminate_macro_nodes---------------------- 2414 // Eliminate scalar replaced allocations and associated locks. 2415 void PhaseMacroExpand::eliminate_macro_nodes() { 2416 if (C->macro_count() == 0) 2417 return; 2418 2419 // First, attempt to eliminate locks 2420 int cnt = C->macro_count(); 2421 for (int i=0; i < cnt; i++) { 2422 Node *n = C->macro_node(i); 2423 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2424 // Before elimination mark all associated (same box and obj) 2425 // lock and unlock nodes. 2426 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2427 } 2428 } 2429 bool progress = true; 2430 while (progress) { 2431 progress = false; 2432 for (int i = C->macro_count(); i > 0; i--) { 2433 Node * n = C->macro_node(i-1); 2434 bool success = false; 2435 debug_only(int old_macro_count = C->macro_count();); 2436 if (n->is_AbstractLock()) { 2437 success = eliminate_locking_node(n->as_AbstractLock()); 2438 } 2439 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2440 progress = progress || success; 2441 } 2442 } 2443 // Next, attempt to eliminate allocations 2444 _has_locks = false; 2445 progress = true; 2446 while (progress) { 2447 progress = false; 2448 for (int i = C->macro_count(); i > 0; i--) { 2449 Node * n = C->macro_node(i-1); 2450 bool success = false; 2451 debug_only(int old_macro_count = C->macro_count();); 2452 switch (n->class_id()) { 2453 case Node::Class_Allocate: 2454 case Node::Class_AllocateArray: 2455 success = eliminate_allocate_node(n->as_Allocate()); 2456 break; 2457 case Node::Class_CallStaticJava: 2458 success = eliminate_boxing_node(n->as_CallStaticJava()); 2459 break; 2460 case Node::Class_Lock: 2461 case Node::Class_Unlock: 2462 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2463 _has_locks = true; 2464 break; 2465 default: 2466 assert(n->Opcode() == Op_LoopLimit || 2467 n->Opcode() == Op_Opaque1 || 2468 n->Opcode() == Op_Opaque2 || 2469 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2470 } 2471 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2472 progress = progress || success; 2473 } 2474 } 2475 } 2476 2477 //------------------------------expand_macro_nodes---------------------- 2478 // Returns true if a failure occurred. 2479 bool PhaseMacroExpand::expand_macro_nodes() { 2480 // Last attempt to eliminate macro nodes. 2481 eliminate_macro_nodes(); 2482 2483 // Make sure expansion will not cause node limit to be exceeded. 2484 // Worst case is a macro node gets expanded into about 50 nodes. 2485 // Allow 50% more for optimization. 2486 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 2487 return true; 2488 2489 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2490 bool progress = true; 2491 while (progress) { 2492 progress = false; 2493 for (int i = C->macro_count(); i > 0; i--) { 2494 Node * n = C->macro_node(i-1); 2495 bool success = false; 2496 debug_only(int old_macro_count = C->macro_count();); 2497 if (n->Opcode() == Op_LoopLimit) { 2498 // Remove it from macro list and put on IGVN worklist to optimize. 2499 C->remove_macro_node(n); 2500 _igvn._worklist.push(n); 2501 success = true; 2502 } else if (n->Opcode() == Op_CallStaticJava) { 2503 // Remove it from macro list and put on IGVN worklist to optimize. 2504 C->remove_macro_node(n); 2505 _igvn._worklist.push(n); 2506 success = true; 2507 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2508 _igvn.replace_node(n, n->in(1)); 2509 success = true; 2510 #if INCLUDE_RTM_OPT 2511 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2512 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2513 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2514 Node* cmp = n->unique_out(); 2515 #ifdef ASSERT 2516 // Validate graph. 2517 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2518 BoolNode* bol = cmp->unique_out()->as_Bool(); 2519 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2520 (bol->_test._test == BoolTest::ne), ""); 2521 IfNode* ifn = bol->unique_out()->as_If(); 2522 assert((ifn->outcnt() == 2) && 2523 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change), ""); 2524 #endif 2525 Node* repl = n->in(1); 2526 if (!_has_locks) { 2527 // Remove RTM state check if there are no locks in the code. 2528 // Replace input to compare the same value. 2529 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2530 } 2531 _igvn.replace_node(n, repl); 2532 success = true; 2533 #endif 2534 } 2535 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2536 progress = progress || success; 2537 } 2538 } 2539 2540 // expand "macro" nodes 2541 // nodes are removed from the macro list as they are processed 2542 while (C->macro_count() > 0) { 2543 int macro_count = C->macro_count(); 2544 Node * n = C->macro_node(macro_count-1); 2545 assert(n->is_macro(), "only macro nodes expected here"); 2546 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2547 // node is unreachable, so don't try to expand it 2548 C->remove_macro_node(n); 2549 continue; 2550 } 2551 switch (n->class_id()) { 2552 case Node::Class_Allocate: 2553 expand_allocate(n->as_Allocate()); 2554 break; 2555 case Node::Class_AllocateArray: 2556 expand_allocate_array(n->as_AllocateArray()); 2557 break; 2558 case Node::Class_Lock: 2559 expand_lock_node(n->as_Lock()); 2560 break; 2561 case Node::Class_Unlock: 2562 expand_unlock_node(n->as_Unlock()); 2563 break; 2564 default: 2565 assert(false, "unknown node type in macro list"); 2566 } 2567 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2568 if (C->failing()) return true; 2569 } 2570 2571 _igvn.set_delay_transform(false); 2572 _igvn.optimize(); 2573 if (C->failing()) return true; 2574 return false; 2575 }